Complete-cover prediction comparison from saved certificates

Use only saved earlier and direct-source reports under the qualified reuse adapter. The original measurements, six model/processing cases, eight regions per case, 51 expression columns, 78 enclosing coordinates, observation allowance, arithmetic margin and original twofold-width target remain unchanged.

Verify provider qualification, failed-task reasons, source and artifact hashes, numerical environments, common maps, domains and conditioning. A completed report saved before a later time limit is distinct from a completed source task. Keep every source study's failure label unchanged. Reject reports implicated in a mathematical or known-truth failure.

Complete coverage rule

Inventory all 48 regions before launch, selecting providers by compatible available certificates, without consulting future truth or model costs. Five groups have all required saved queries, permitting 40 fresh region reconstructions. The zero-field globe with profiled removal lacks future reports in two regions; use only its unchanged complete inherited group range. Do not use its partial new union. Preserve its eight inventory rows with that disposition.

For each fresh region select the minimum valid coordinate remainder, rebuild training constraints, preserve inherited rows, verify the parent support and child domain, contract with the existing eight-pass limit, and independently audit the six future endpoints. Check known truth afterward. Every scheduled region must pass; partial successful subsets cannot qualify the new method.

Verify that the inherited complete 192-box comparison descends from the same parent cases and coordinate frames. Union all surviving regions in a complete fresh group, then intersect only complete valid group ranges. Publish fresh coverage, inherited coverage, final ranges and baseline-only disposition separately. All six final complete covers must meet the original target.

Separate finite allowance

• 40 fresh-region identities; 280 saved-report merge calls, zero propagation steps.
• At most 560 provider report reads, with the exact counts frozen per region; 160 provider manifest/receipt reads.
• 40 reference setups and child proofs; 240 parent support verifications.
• 1,440 training-row audits, 40 contractions and 40 truth audits.
• At most 240 one-second LP proposals and 960 helper/independent future support checks, in addition to the 240 parent checks.
• Four workers, one numerical thread each; four total threads. 600 seconds overall, checked cooperatively. No retries, signals or allowance expansion.
• Zero native trajectory calls, integration packet steps, compiler invocations, optimizer starts, original-aircraft reads or protected ESA data.

Freeze source, exact provider availability, input/provenance hashes, environment, frames and acceptance rules before launch. This does not relabel either earlier study or qualify the direct-only method. Passing would establish the utility of the combined certificates on the existing generated example, not recording-scale adequacy, observational confidence or an additional empirical model preference.

Adapter qualification · Earlier direct-only outcome · Roadmap.